Exploratory Training for Universal Lesion Detection: Enhancing Lesion Mining Quality Through Temporal Verification.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Universal lesion detection (ULD) is crucial for clinical practice, enabling the identification of diverse lesions across multiple organs.
• Deep learning models for ULD require high-quality annotated data, which is often incomplete in real-world scenarios due to cost, expertise, and lesion variability.
• Existing pseudo-labeling methods use mini-batch level lesion mining, but the inconsistent quality of mined lesions across iterations limits performance.

Purpose of the Study:

• To develop an improved method for training universal lesion detection (ULD) models using incomplete annotations.
• To address the limitations of current pseudo-labeling techniques by enhancing the reliability of mined lesions.
• To introduce a novel exploratory-training-based ULD (ET-ULD) approach that leverages temporal consistency for lesion selection.

Main Methods:

• Proposed an exploratory-training-based ULD (ET-ULD) method employing a teacher-student detection model architecture.
• The teacher model mines suspicious lesions, which are then combined with incomplete annotations to train the student model.
• Implemented a bounding-box bank to record lesion mining timestamps across multiple training rounds, enabling reliability assessment based on consistent appearance over time.

Main Results:

• ET-ULD demonstrated superior performance compared to existing state-of-the-art methods on two distinct lesion image datasets.
• The method achieved a significant 5.4% improvement in Average Precision (AP) on the DeepLesion dataset.
• Experimental results validated the effectiveness of assessing mined lesion reliability through consistent temporal appearance.

Conclusions:

• The proposed ET-ULD method effectively improves ULD performance when trained with incomplete annotations by reliably selecting pseudo-labels.
• Assessing lesion reliability over time is a crucial criterion for selecting high-quality pseudo-labels, surpassing dynamic mini-batch mining.
• ET-ULD offers a promising solution for developing robust deep learning-based lesion detection systems in clinical settings.